Skip to content

Commit f560144

Browse files
overmightyovermighty
committed
Move both as() implementations to public member functions
1 parent 0159939 commit f560144

File tree

1 file changed

+123
-118
lines changed

1 file changed

+123
-118
lines changed

‎libc/src/__support/FPUtil/dyadic_float.h

Lines changed: 123 additions & 118 deletions
Original file line numberDiff line numberDiff line change
@@ -37,8 +37,7 @@ namespace fputil {
3737
// The outputs of the constructors and most functions will be normalized.
3838
// To simplify and improve the efficiency, many functions will assume that the
3939
// inputs are normal.
40-
template <size_t Bits> class DyadicFloat {
41-
public:
40+
template <size_t Bits> struct DyadicFloat {
4241
using MantissaType = LIBC_NAMESPACE::UInt<Bits>;
4342

4443
Sign sign = Sign::POS;
@@ -102,18 +101,116 @@ template <size_t Bits> class DyadicFloat {
102101
return exponent + (Bits - 1);
103102
}
104103

105-
// Assume that it is already normalized.
106-
// Output is rounded correctly with respect to the current rounding mode.
104+
template <typename T>
105+
LIBC_INLINE constexpr cpp::enable_if_t<
106+
cpp::is_floating_point_v<T> && (FPBits<T>::FRACTION_LEN < Bits), T>
107+
generic_as() const {
108+
using FPBits = FPBits<float16>;
109+
using StorageType = typename FPBits::StorageType;
110+
111+
constexpr int EXTRA_FRACTION_LEN = Bits - 1 - FPBits::FRACTION_LEN;
112+
113+
if (mantissa == 0)
114+
return FPBits::zero(sign).get_val();
115+
116+
int unbiased_exp = get_unbiased_exponent();
117+
118+
if (unbiased_exp + FPBits::EXP_BIAS >= FPBits::MAX_BIASED_EXPONENT) {
119+
set_errno_if_required(ERANGE);
120+
raise_except_if_required(FE_OVERFLOW | FE_INEXACT);
121+
122+
switch (quick_get_round()) {
123+
case FE_TONEAREST:
124+
return FPBits::inf(sign).get_val();
125+
case FE_TOWARDZERO:
126+
return FPBits::max_normal(sign).get_val();
127+
case FE_DOWNWARD:
128+
if (sign.is_pos())
129+
return FPBits::max_normal(Sign::POS).get_val();
130+
return FPBits::inf(Sign::NEG).get_val();
131+
case FE_UPWARD:
132+
if (sign.is_neg())
133+
return FPBits::max_normal(Sign::NEG).get_val();
134+
return FPBits::inf(Sign::POS).get_val();
135+
default:
136+
__builtin_unreachable();
137+
}
138+
}
139+
140+
StorageType out_biased_exp = 0;
141+
StorageType out_mantissa = 0;
142+
bool round = false;
143+
bool sticky = false;
144+
bool underflow = false;
145+
146+
if (unbiased_exp < -FPBits::EXP_BIAS - FPBits::FRACTION_LEN) {
147+
sticky = true;
148+
underflow = true;
149+
} else if (unbiased_exp == -FPBits::EXP_BIAS - FPBits::FRACTION_LEN) {
150+
round = true;
151+
MantissaType sticky_mask = (MantissaType(1) << (Bits - 1)) - 1;
152+
sticky = (mantissa & sticky_mask) != 0;
153+
} else {
154+
int extra_fraction_len = EXTRA_FRACTION_LEN;
155+
156+
if (unbiased_exp < 1 - FPBits::EXP_BIAS) {
157+
underflow = true;
158+
extra_fraction_len += 1 - FPBits::EXP_BIAS - unbiased_exp;
159+
} else {
160+
out_biased_exp =
161+
static_cast<StorageType>(unbiased_exp + FPBits::EXP_BIAS);
162+
}
163+
164+
MantissaType round_mask = MantissaType(1) << (extra_fraction_len - 1);
165+
round = (mantissa & round_mask) != 0;
166+
MantissaType sticky_mask = round_mask - 1;
167+
sticky = (mantissa & sticky_mask) != 0;
168+
169+
out_mantissa = static_cast<StorageType>(mantissa >> extra_fraction_len);
170+
}
171+
172+
bool lsb = (out_mantissa & 1) != 0;
173+
174+
StorageType result =
175+
FPBits::create_value(sign, out_biased_exp, out_mantissa).uintval();
176+
177+
switch (quick_get_round()) {
178+
case FE_TONEAREST:
179+
if (round && (lsb || sticky))
180+
++result;
181+
break;
182+
case FE_DOWNWARD:
183+
if (sign.is_neg() && (round || sticky))
184+
++result;
185+
break;
186+
case FE_UPWARD:
187+
if (sign.is_pos() && (round || sticky))
188+
++result;
189+
break;
190+
default:
191+
break;
192+
}
193+
194+
if (round || sticky) {
195+
int excepts = FE_INEXACT;
196+
if (FPBits(result).is_inf()) {
197+
set_errno_if_required(ERANGE);
198+
excepts |= FE_OVERFLOW;
199+
} else if (underflow) {
200+
set_errno_if_required(ERANGE);
201+
excepts |= FE_UNDERFLOW;
202+
}
203+
raise_except_if_required(excepts);
204+
}
205+
206+
return FPBits(result).get_val();
207+
}
208+
107209
template <typename T, bool ShouldSignalExceptions,
108210
typename = cpp::enable_if_t<cpp::is_floating_point_v<T> &&
109211
(FPBits<T>::FRACTION_LEN < Bits),
110212
void>>
111-
LIBC_INLINE constexpr T as() const {
112-
#if defined(LIBC_TYPES_HAS_FLOAT16) && !defined(__LIBC_USE_FLOAT16_CONVERSION)
113-
if constexpr (cpp::is_same_v<T, float16>)
114-
return generic_as<T>();
115-
#endif
116-
213+
LIBC_INLINE constexpr T fast_as() const {
117214
if (LIBC_UNLIKELY(mantissa.is_zero()))
118215
return FPBits<T>::zero(sign).get_val();
119216

@@ -234,6 +331,22 @@ template <size_t Bits> class DyadicFloat {
234331
return r;
235332
}
236333

334+
// Assume that it is already normalized.
335+
// Output is rounded correctly with respect to the current rounding mode.
336+
template <typename T, bool ShouldSignalExceptions,
337+
typename = cpp::enable_if_t<cpp::is_floating_point_v<T> &&
338+
(FPBits<T>::FRACTION_LEN < Bits),
339+
void>>
340+
LIBC_INLINE constexpr T as() const {
341+
#if defined(LIBC_TYPES_HAS_FLOAT16) && !defined(__LIBC_USE_FLOAT16_CONVERSION)
342+
if constexpr (cpp::is_same_v<T, float16>) {
343+
static_assert(ShouldSignalExceptions);
344+
return generic_as<T>();
345+
}
346+
#endif
347+
return fast_as<T, ShouldSignalExceptions>();
348+
}
349+
237350
template <typename T,
238351
typename = cpp::enable_if_t<cpp::is_floating_point_v<T> &&
239352
(FPBits<T>::FRACTION_LEN < Bits),
@@ -259,114 +372,6 @@ template <size_t Bits> class DyadicFloat {
259372

260373
return new_mant;
261374
}
262-
263-
private:
264-
template <typename OutType>
265-
LIBC_INLINE constexpr cpp::enable_if_t<
266-
cpp::is_floating_point_v<OutType> &&
267-
sizeof(typename FPBits<OutType>::StorageType) <= sizeof(MantissaType),
268-
OutType>
269-
generic_as() const {
270-
using FPBits = FPBits<float16>;
271-
using StorageType = typename FPBits::StorageType;
272-
273-
constexpr int EXTRA_FRACTION_LEN = Bits - 1 - FPBits::FRACTION_LEN;
274-
275-
if (mantissa == 0)
276-
return FPBits::zero(sign).get_val();
277-
278-
int unbiased_exp = get_unbiased_exponent();
279-
280-
if (unbiased_exp + FPBits::EXP_BIAS >= FPBits::MAX_BIASED_EXPONENT) {
281-
set_errno_if_required(ERANGE);
282-
raise_except_if_required(FE_OVERFLOW | FE_INEXACT);
283-
284-
switch (quick_get_round()) {
285-
case FE_TONEAREST:
286-
return FPBits::inf(sign).get_val();
287-
case FE_TOWARDZERO:
288-
return FPBits::max_normal(sign).get_val();
289-
case FE_DOWNWARD:
290-
if (sign.is_pos())
291-
return FPBits::max_normal(Sign::POS).get_val();
292-
return FPBits::inf(Sign::NEG).get_val();
293-
case FE_UPWARD:
294-
if (sign.is_neg())
295-
return FPBits::max_normal(Sign::NEG).get_val();
296-
return FPBits::inf(Sign::POS).get_val();
297-
default:
298-
__builtin_unreachable();
299-
}
300-
}
301-
302-
StorageType out_biased_exp = 0;
303-
StorageType out_mantissa = 0;
304-
bool round = false;
305-
bool sticky = false;
306-
bool underflow = false;
307-
308-
if (unbiased_exp < -FPBits::EXP_BIAS - FPBits::FRACTION_LEN) {
309-
sticky = true;
310-
underflow = true;
311-
} else if (unbiased_exp == -FPBits::EXP_BIAS - FPBits::FRACTION_LEN) {
312-
round = true;
313-
MantissaType sticky_mask = (MantissaType(1) << (Bits - 1)) - 1;
314-
sticky = (mantissa & sticky_mask) != 0;
315-
} else {
316-
int extra_fraction_len = EXTRA_FRACTION_LEN;
317-
318-
if (unbiased_exp < 1 - FPBits::EXP_BIAS) {
319-
underflow = true;
320-
extra_fraction_len += 1 - FPBits::EXP_BIAS - unbiased_exp;
321-
} else {
322-
out_biased_exp =
323-
static_cast<StorageType>(unbiased_exp + FPBits::EXP_BIAS);
324-
}
325-
326-
MantissaType round_mask = MantissaType(1) << (extra_fraction_len - 1);
327-
round = (mantissa & round_mask) != 0;
328-
MantissaType sticky_mask = round_mask - 1;
329-
sticky = (mantissa & sticky_mask) != 0;
330-
331-
out_mantissa = static_cast<StorageType>(mantissa >> extra_fraction_len);
332-
}
333-
334-
bool lsb = (out_mantissa & 1) != 0;
335-
336-
StorageType result =
337-
FPBits::create_value(sign, out_biased_exp, out_mantissa).uintval();
338-
339-
switch (quick_get_round()) {
340-
case FE_TONEAREST:
341-
if (round && (lsb || sticky))
342-
++result;
343-
break;
344-
case FE_DOWNWARD:
345-
if (sign.is_neg() && (round || sticky))
346-
++result;
347-
break;
348-
case FE_UPWARD:
349-
if (sign.is_pos() && (round || sticky))
350-
++result;
351-
break;
352-
default:
353-
break;
354-
}
355-
356-
if (round || sticky) {
357-
int excepts = FE_INEXACT;
358-
if (FPBits(result).is_inf()) {
359-
set_errno_if_required(ERANGE);
360-
excepts |= FE_OVERFLOW;
361-
} else if (underflow) {
362-
set_errno_if_required(ERANGE);
363-
excepts |= FE_UNDERFLOW;
364-
}
365-
raise_except_if_required(excepts);
366-
}
367-
368-
return FPBits(result).get_val();
369-
}
370375
};
371376

372377
// Quick add - Add 2 dyadic floats with rounding toward 0 and then normalize the

0 commit comments

Comments
 (0)